Time lag prediction method and device for SCR denitration system in thermal power plant

Abstract

The present invention provides a method and device for predicting the time lag of a SCR denitration system in a thermal power plant. The method comprises: selecting the SCR inlet NO according to the parameter-related selection principle within a set time period. x Concentration prediction parameters; Based on the prediction parameters, establish the SCR inlet NO x Concentration prediction model; Based on the time-lag shape difference rate method, a time-lag prediction model for the valve opening of the denitrification device is established; Based on the SCR inlet NO x The concentration prediction model and the time lag prediction model predict the time lag of the ammonia production system and obtain the prediction results. The prediction method of the present invention realizes the prediction of the time lag of the ammonia production system and can accurately predict the magnitude of the ammonia production time lag so that the ammonia injection system action and the SCR inlet NO x Content matching improves the prediction accuracy, saves costs for the overall operation of the SCR denitrification system, and provides suggestions for optimizing the control of the valve opening of the denitrification device.

Description

Technical Field

[0001] The invention belongs to the technical field of thermal power denitration, and in particular relates to a method and a device for predicting the time lag of a SCR denitration system in a thermal power plant. Background Art

[0002] There are currently two main technical routes for removing nitrogen oxides from flue gas in coal-fired power plants. One is the in-furnace low-nitrogen combustion technology that reduces nitrogen oxides produced during boiler combustion, and the other is the flue gas denitrification technology that decomposes the nitrogen oxides that have already been generated. The main technologies for flue gas denitrification include gas phase reaction method, liquid absorption method, adsorption method, liquid membrane method, microbial method, etc. Among the above-mentioned flue gas denitrification technologies, selective catalytic reduction technology (SCR) is the most widely used due to its dual advantages of high denitrification efficiency and mature technology. There are more than 120 large-scale SCR devices in operation in Europe, which can achieve more than 80% NO x Removal efficiency: Since the SCR system was first used in the Shimoneski thermal power plant in Japan in 1975, SCR technology has become the most mainstream flue gas denitrification technology in the world.

[0003] There are generally three options for the reducing agent required in the SCR reaction, namely pure ammonia, ammonia water and urea, corresponding to three different ammonia production processes. Due to the time lag in the production of ammonia from urea, the response between the ammonia production system and the ammonia injection device cannot be synchronized with the real-time signal (SCR inlet NO x The maximum time lag between the two can reach 120 seconds, but in most cases 120s is greater than the SCR inlet NO x The half-period of fluctuation of the content of ammonia-NO x The matching error problem is very serious. Therefore, thermal power plants need to increase the amount of ammonia injection to meet the denitrification effect. Excessive ammonia injection will lead to a waste of raw materials, increased ammonia leakage, affect the operation of the air preheater, and aggravate the catalyst passivation in the SCR reactor, which will eventually lead to an increase in the overall operating cost of the flue gas denitrification system. Due to the influence of the ammonia production system mechanism, the ammonia production time lag varies with the SCR inlet NO x In order to accurately predict the time lag magnitude so that the ammonia injection system can operate and the SCR inlet NO x In order to match the content, it is necessary to establish a prediction model to describe the time lag of the ammonia injection system.

[0004] Since the historical operation data provided by the thermal power plant does not contain time-lag values, only the ammonia distribution valve opening data is provided to indirectly reflect the time-lag values. Although the time-lag data can be obtained through manual calibration, the manual identification efficiency is low and cannot meet the data volume requirements. Therefore, it is impossible to establish a time-lag prediction model starting from the flue gas duct through a data-driven method. At the same time, the historical operation data does not include the relevant parameters of the ammonia production system, and the thermal power plant has not provided a detailed design of the ammonia production system. Therefore, it is impossible to start from the ammonia production system data to carry out data-driven time-lag modeling or mechanism modeling.

[0005] In view of the above problems, it is necessary to propose a method and device for predicting the time lag of the SCR denitrification system in a thermal power plant that is reasonably designed and effectively solves the above problems. Summary of the invention

[0006] The present invention aims to solve at least one of the technical problems existing in the prior art and provides a method and device for predicting the time lag of an SCR denitration system in a thermal power plant.

[0007] One aspect of the present invention provides a method for predicting the time lag of a SCR denitration system in a thermal power plant, the method comprising:

[0008] During the set time period, select SCR inlet NO according to the parameter selection principle. x Concentration prediction parameters;

[0009] Based on the prediction parameters, the SCR inlet NO is established. x Concentration prediction model;

[0010] Based on the time-lag shape difference rate method, a time-lag prediction model for the valve opening of the denitration device is established;

[0011] Based on the SCR inlet NO x The concentration prediction model and the time lag prediction model predict the time lag of the ammonia production system and obtain prediction results.

[0012] Optionally, the time-lag prediction model of the valve opening of the denitration device is established based on the time-lag shape difference rate method, including:

[0013] The time-delay shape difference rate method is used to determine the SCR inlet NO x The concentration and its rate of change are input, and the time-lag value is output. The prediction model of the time-lag value of the valve opening of the denitrification device is obtained by segmented fitting;

[0014] The time-delay shape error rate method is:

[0015]

[0016] Where T is the offset time δT, and the partial curve shape difference rate when the offset time is T, t1 and t2 represent the partial SCR inlet NO x The starting and ending time points of the concentration curve, ρ and C represent the SCR inlet NO x Concentration and ammonia regulating valve opening, k t (ρ) and k t (C) respectively represent SCR inlet NO x The backward change rate of concentration and ammonia regulating valve opening at time t.

[0017] Optionally, the SCR inlet NO is established based on the prediction parameters. x Concentration prediction models, including:

[0018] Based on the prediction parameters, obtaining corresponding prediction parameter data at each moment;

[0019] Based on the prediction parameter data corresponding to each moment, obtaining time series data;

[0020] Based on the time series data, the SCR inlet NO is established. x Concentration prediction model.

[0021] Optionally, the SCR inlet NO is established based on the time series data. x Concentration prediction models, including:

[0022] Take the time series data as the SCR inlet NO x The input data of the concentration prediction model, and in the process of establishing the prediction model, the hidden state of the prediction model is updated; wherein,

[0023] The expression of the hidden state is:

[0024] h t =f(h t-1 , x t )=f(Uh t-1 +Wx t +b),

[0025] Among them, U∈R G*G , W∈R G*k , b∈R G are model parameters, G is the hidden layer dimension, k is the input dimension, and f is the tanh activation function.

[0026] Optionally, the SCR inlet NO is established based on the time series data. x After the concentration prediction model is obtained, the method further comprises:

[0027] Based on the SCR inlet NOx The concentration prediction model is trained, wherein when the training residual is less than a preset threshold, the SCR inlet NO x Concentration prediction model.

[0028] Optionally, the SCR inlet NO x The concentration prediction model and the time lag prediction model predict the time lag of the ammonia production system and obtain prediction results, including:

[0029] Through SCR inlet NO x The concentration time series prediction model predicts the SCR inlet NO in a preset time period. x The content fluctuation data is used to predict the SCR inlet NO x Time series prediction curve of content;

[0030] Adopt SCR inlet NO x The content hysteresis prediction model for the SCR inlet NO x The time series prediction value at any time on the time series prediction curve of the content is predicted to obtain the SCR inlet NO x Time-hysteresis numerical curve corresponding to the content;

[0031] According to the timing prediction curve and the time hysteresis value curve, the SCR inlet NO x Content tracking curve; among them,

[0032] The SCR inlet NO x The content tracking curve is:

[0033] ρ trace (t) = ρ(t + D(d)),

[0034] Among them, ρ(d) is the time series prediction curve, d is the difference between any point on the time series prediction curve and time t0 in the time direction, D(d) is the time hysteresis numerical curve, ρ trace (t) is the time deviation d corresponding to any tracking moment on the timing prediction curve ρ(d) of the SCR inlet NO x Predicted content value;

[0035] The SCR inlet tracking NO at the current moment x The content is the current SCR inlet NO x The content signal is transmitted to the flue gas denitrification system as an input signal to obtain a time-delay prediction result.

[0036] Optionally, the SCR entry NO is established based on the LSTM neural network algorithm. x Concentration prediction model.

[0037] Optionally, the parameter correlation selection principle adopts the principle of maximum correlation and minimum redundancy; wherein,

[0038] The expression of the maximum correlation is:

[0039]

[0040]

[0041] Among them, maxD(S,c) is the maximum correlation of discrete variables, D is a categorical variable, S is a feature subset, c is a categorical variable, I is a feature variable, xi is the i-th feature, maxDF is the maximum correlation of continuous variables, and F(xi,c) is the F statistic;

[0042] The expression of the minimum redundancy is:

[0043]

[0044]

[0045] Among them, minR(S) is the minimum redundancy of discrete variables, minRc is the minimum redundancy of continuous variables, and c(xi, xj) is the correlation function.

[0046] Optionally, the SCR inlet NO x The concentration prediction parameters include SCR inlet flue gas temperature, SCR inlet flue gas oxygen content, load and total air volume.

[0047] Another aspect of the present invention provides a device for predicting the time lag of an SCR denitration system in a thermal power plant, the device comprising:

[0048] The selection module is used to select the SCR inlet NO according to the parameter-related selection principle within the set time period. x Concentration prediction parameters;

[0049] The first establishing module is used to establish the SCR inlet NO based on the prediction parameter. x Concentration prediction model;

[0050] The second establishment module is used to establish a time-lag prediction model for the valve opening of the denitration device based on the time-lag shape difference rate method;

[0051] Prediction module for the SCR inlet NO x The concentration prediction model and the time lag prediction model predict the time lag of the ammonia production system and obtain prediction results.

[0052] The method and device for predicting the time lag of the SCR denitration system of a thermal power plant of the present invention include: selecting the SCR inlet NO in a set time period according to the parameter-related selection principle; x Concentration prediction parameters; Based on the prediction parameters, establish the SCR inlet NO x Concentration prediction model; Based on the time-lag shape difference rate method, a time-lag prediction model for the valve opening of the denitrification device is established; Based on the SCR inlet NO x The concentration prediction model and the time lag prediction model predict the time lag of the ammonia production system and obtain the prediction results. The time lag prediction method of the SCR denitrification system in a thermal power plant of the present invention realizes the prediction of the time lag of the ammonia production system and can accurately predict the magnitude of the ammonia production time lag so that the ammonia injection system action and the SCR inlet NO x Content matching improves the prediction accuracy, saves costs for the overall operation of the SCR denitrification system, and provides suggestions for optimizing the control of the valve opening of the denitrification device. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 A schematic flow chart of a method for predicting time lag of a thermal power plant SCR denitration system according to an embodiment of the present invention;

[0054] Figure 2 In another embodiment of the present invention, a method for predicting the time lag of a thermal power plant SCR denitrification system is used to establish an SCR inlet NO x Schematic diagram of the flow chart of the concentration prediction model;

[0055] Figure 3 This is a schematic diagram of a flow chart of a device for predicting time lag of a SCR denitration system in a thermal power plant according to an embodiment of the present invention;

[0056] Figure 4 FIG. 4 is a schematic diagram of the composition of an electronic device in another embodiment of the present invention. DETAILED DESCRIPTION

[0057] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0058] like Figure 1 As shown, one aspect of the present invention provides a method S100 for predicting the time lag of a thermal power plant SCR denitration system, and the prediction method S100 includes:

[0059] S110, within the set time period, select SCR inlet NO according to the parameter selection principle x Concentration prediction parameters.

[0060] Specifically, first, a time period needs to be selected; then, the SCR inlet NO xConcentration prediction parameters, where the prediction parameters include at least but are not limited to SCR inlet flue gas temperature, SCR inlet flue gas oxygen content, load and total air volume. Finally, within the set time period, based on the principle of maximum relevance and minimum redundancy, the prediction parameters are selected, where:

[0061] The expression for the maximum correlation is:

[0062]

[0063]

[0064] Among them, maxD(S,c) is the maximum correlation of discrete variables, D is a categorical variable, S is a feature subset, c is a categorical variable, I is a feature variable, xi is the i-th feature, maxDF is the maximum correlation of continuous variables, and F(xi,c) is the F statistic;

[0065] The expression of minimum redundancy is:

[0066]

[0067]

[0068] Among them, minR(S) is the minimum redundancy of discrete variables, minRc is the minimum redundancy of continuous variables, and c(xi, xj) is the correlation function.

[0069] The maximum correlation and minimum redundancy are obtained based on mutual information. Mutual information can measure the correlation between two variables. The size of mutual information represents the size of the association between the two variables. By selecting a set number of features with the largest mutual information between the two variables from all features, we can get the set number of features that are most correlated between the two variables.

[0070] The expression of mutual information is:

[0071]

[0072] Among them, x is a variable, y is a variable, and I(x, y) is the mutual information.

[0073] S120: Establish the SCR inlet NO based on the prediction parameters. x Concentration prediction model.

[0074] Establish the SCR inlet NO xThe concentration prediction model appropriately selects the corresponding prediction model to be established based on the characteristics of the selected prediction parameters. The selection of prediction parameters is based on the selection of their features. Feature selection, also known as feature subset selection or attribute selection, refers to the process of selecting certain features from multiple existing features to optimize specific indicators of the system. It is the process of selecting some of the most effective features from the original features to reduce the dimension of the data set. It is not only an important means to improve the performance of the learning algorithm, but also a key data preprocessing step in pattern recognition. For a prediction model, good learning samples are the key to training the prediction model.

[0075] like Figure 2 As shown, S120 specifically includes:

[0076] S121. Based on the prediction parameters, obtain corresponding prediction parameter data at each moment.

[0077] Specifically, among the selected prediction parameters, at a certain moment, a certain prediction parameter will correspond to a specific prediction parameter data, and accordingly, different prediction parameters will correspond to different prediction parameter data. At each moment, there will be several corresponding prediction parameter data.

[0078] For example, at 9:01, the predicted parameter SCR inlet NO x The concentration is 350 mg / m3, and the predicted parameters SCR inlet NO at 9:03 x If the concentration is 300 mg per cubic meter, then there will be some predicted parameter data within the 2 minutes from 9:01 to 9:03.

[0079] The selected prediction data is preprocessed to obtain the required prediction parameter data, wherein the preprocessing method includes at least but is not limited to data singular point removal, normalization, standardization, filtering, noise reduction, feature decomposition, and time-frequency domain analysis.

[0080] For example, the selected prediction data is preprocessed by singular point removal, and a smoothing program is written: the data is read in sequence, and some continuous data are saved as variables. The relationship between a certain data and the data before and after it is determined. If the difference is too large, the current value is replaced with the median or average value of the continuous data. If it exceeds the pre-set threshold of the singular point, it is removed. The above threshold needs to be visually inspected to determine a suitable value. This method can effectively remove data points that do not meet the standards.

[0081] In another exemplary embodiment, the selected prediction data is preprocessed by normalization. There are two forms of normalization methods, one is to convert the number into a decimal between (0, 1), and the other is to convert the dimensioned expression into a dimensionless expression. It is mainly proposed for the convenience of data processing. Mapping the data to the range of 0 to 1 for processing is more convenient and fast, and the data to be processed is classified into the category of digital signal processing. Then, the selected prediction data is preprocessed by normalization. For example, the range of the selected prediction data is divided into four categories. For example, the range of the selected prediction data is {0, 1}, and the range {0, 1} is divided into four categories: 0-1 / 4, 1 / 4-1 / 2, 1 / 2-3 / 4, 3 / 4-1, and the four categories are marked with categories, and the data is obtained.

[0082] S122. Obtain time series data based on the corresponding prediction parameter data at each moment.

[0083] Specifically, a certain prediction parameter data at a certain moment and the prediction parameter data at the previous moment, the previous moment, the next moment, the next moment, etc. constitute a certain time series data. Different prediction parameter data at each moment correspondingly constitute a series of time series data with the prediction parameter data at the previous moment, the previous moment, the next moment, the next moment, etc. of each moment.

[0084] S123, based on the time series data, establish the SCR inlet NO x Concentration prediction model.

[0085] SCR inlet NO x The establishment of a concentration prediction model requires input data, and time series data is used as input data. In the process of establishing a prediction model, the hidden state in the prediction model is continuously updated. The continuous updating of the hidden state can reflect the characteristics existing in the time series data. Since the establishment of the prediction model requires the characteristics of the time series data for feature selection, the above reflection of the characteristics existing in the time series data paves the way for the establishment of the prediction model. Among them, the hidden state is the invisible state.

[0086] The expression of the hidden state is:

[0087] h t =f(h t-1 , x t )=f(Uh t-1 +Wx t +b),

[0088] Among them, U∈R G*G , W∈R G*k , b∈R Gare model parameters, G is the hidden layer dimension, k is the input dimension, and f is the tanh activation function.

[0089] For example, LSTM (Long Short-Term Memory) is a time recurrent neural network. To build the above model, it is necessary to first define RNN (Recurrent Neural Network). The obtained time series data is input into RNN, and the corresponding input of the time series data at each moment is given to each repetitive unit in the RNN result to continuously update the hidden state. Since the LSTM algorithm contains memory units, the memory units will memorize the corresponding hidden states. The update of the hidden state and the memory unit requires three gates: forget gate, input gate, and output gate to control, so as to protect the control unit state.

[0090] Among them, the forget gate needs to forget or discard some information. The task of the forget gate is to accept a long-term memory (the output from the previous unit module) and decide which part to keep and which part to forget. For example, in a language model, we want to predict the next word from a word. The unit module may contain a certain attribute that is not needed in the next module unit, so this attribute can be forgotten or discarded in the unit module.

[0091] Memory gate, the function of the memory gate is to determine what kind of new information is stored in the unit module. The memory gate consists of two parts: the sigmoid layer and the tanh layer. The sigmoid layer determines what value needs to be updated; the tanh layer creates a new candidate value vector and generates a candidate memory. For the attribute information discarded in the forget gate, find the corresponding new attribute information in this unit module and add it to supplement the discarded attribute information. Among them, the sigmoid layer is an S-type function commonly seen in biology, and the tanh layer is a very simple structure of this repeated structural module in the RNN.

[0092] The output gate is where we finally determine the output value based on the unit module. First, we use a function to determine which part of the unit module needs to be output, and then process the unit module through the tanh layer. Based on the above part that needs to be output and the above processing, we get the information we want to output.

[0093] The disclosed embodiment selects the three gates based on their respective characteristics, which not only enables the LSTM algorithm to delete or add information to the unit state, but also can update the memory unit and the hidden state, all of which are regulated and controlled by the structure of the three gates.

[0094] S124, based on the SCR inlet NO x The concentration prediction model is trained to obtain the SCR inlet NOx Concentration prediction model.

[0095] Specifically, training is performed based on the prediction model, and when the training residual is less than a preset threshold, the target prediction model is obtained. The embodiment of the present disclosure adopts the above method to perform training to obtain the target prediction model, which is simple to operate and effective.

[0096] For example, at the SCR inlet NO x During the training of the concentration prediction model, a threshold is set in advance. When the comprehensive value of each prediction parameter data obtained based on the relevant rules, that is, the training residual, is less than the preset threshold, it is considered that the SCR inlet NO x The concentration prediction model has been successfully trained to obtain the SCR inlet NO x Concentration prediction model. In mathematical statistics, residual refers to the difference between the actual observed value and the estimated value, which contains important information about the basic assumptions of the model. Therefore, the embodiment of the present disclosure selects training residuals to calculate the SCR inlet NO x The criteria for judging whether the concentration prediction model has completed training are more convincing.

[0097] S130. Based on the time-lag shape difference rate method, a time-lag prediction model for the valve opening of the denitration device is established.

[0098] Specifically, the time-delay shape difference rate method is used to calculate the SCR inlet NO x The concentration and its rate of change are input, and the time-lag value is output. The prediction model of the time-lag value of the valve opening of the denitrification device is obtained by segmented fitting;

[0099] The time-delayed shape difference method is:

[0100]

[0101] Where T is the offset time δT, and the partial curve shape difference rate when the offset time is T, t1 and t2 represent the partial SCR inlet NO x The starting and ending time points of the concentration curve, ρ and C represent the SCR inlet NO x Concentration and ammonia regulating valve opening, k t (ρ) and k t (C) respectively represent SCR inlet NO x The backward change rate of concentration and ammonia regulating valve opening at time t.

[0102] SCR inlet NO xThe concentration and its rate of change are input, and the time hysteresis value is output. The prediction model of the time hysteresis value of the valve opening is obtained by segment fitting. The time hysteresis prediction model is divided into three segment functions for implementation. Combined with a large amount of data, the linear correlation is analyzed to establish the NO inlet of SCR. x The time-lag prediction model of the process of a small increase in content was established by adding a deviation coefficient obtained through data analysis and fitting to this model. x The time-lag prediction model for the process of a sharp increase in content was finally selected, taking into account the conditions of over-discharge risk and time-lag fluctuation stability. x The time-lag prediction model for the content reduction process was selected as a fixed value obtained through data analysis. Finally, the SCR inlet NO was completed by verifying the time-lag prediction model on the artificial calibration data. x Content hysteresis prediction model.

[0103] S140, based on the SCR inlet NO x The concentration prediction model and the time lag prediction model predict the time lag of the ammonia production system and obtain prediction results.

[0104] First, through the SCR inlet NO x The concentration time series prediction model predicts the SCR inlet NO in a preset time period. x The content fluctuation data is used to predict the SCR inlet NO x Time series prediction curve of content.

[0105] Specifically, for the SCR inlet NO x The data of a certain moment (called time t0) in the content fluctuation data is obtained by using the SCR inlet NO x The content time series prediction model predicts the data for a period of time after this moment and obtains the time series prediction curve ρ(d), where d is the difference between any point on the time series prediction curve and the moment t0 in the time direction, that is, the time deviation of a certain point, in seconds.

[0106] Secondly, the SCR inlet NO x The content hysteresis prediction model for the SCR inlet NO x The time series prediction value at any time on the time series prediction curve of the content is predicted to obtain the SCR inlet NO x The time-hysteresis numerical curve corresponding to the content.

[0107] Specifically, the predicted value ρ(d0) at any time d0 on the time series prediction curve ρ(d) is calculated using the SCR inlet NO x The content time-lag prediction model obtains the SCR inlet NO at this time point. xThe hysteresis prediction value D(d0) corresponding to the content is obtained by calculating the hysteresis prediction value of each point on the time series prediction curve ρ(d) and connecting the discrete hysteresis values ​​to obtain the hysteresis value curve D(d).

[0108] Here, according to the timing prediction curve and the time hysteresis value curve, the SCR inlet NO x Content tracking curve; among them,

[0109] The SCR inlet NO x The content tracking curve is:

[0110] ρ trace (t) = ρ(t + D(d)),

[0111] Among them, ρ(d) is the time series prediction curve, d is the difference between any point on the time series prediction curve and time t0 in the time direction, D(d) is the time hysteresis numerical curve, ρ trace (t) is the time deviation d corresponding to any tracking moment on the timing prediction curve ρ(d) of the SCR inlet NO x Predicted value of content.

[0112] Specifically, the tracking time t(d) is defined as the time corresponding to the time lag value corresponding to a certain time on the time series prediction model migrating to the negative direction of the time axis, that is, t(d) = t0-D(d). x Content trace (t) is the time deviation d corresponding to any tracking moment on the timing prediction curve ρ(d) of the SCR inlet NO x The content prediction value. For each point on the time series prediction curve, its tracking NO x The SCR inlet NO can be obtained by connecting the discrete results. x Content tracking curve:

[0113] ρ trace (t) = ρ(t + D(d)),

[0114] Among them, ρ(d) is the time series prediction curve, d is the difference between any point on the time series prediction curve and time t0 in the time direction, D(d) is the time hysteresis numerical curve, ρ trace (t) is the time deviation d corresponding to any tracking moment on the timing prediction curve ρ(d) of the SCR inlet NO x Predicted value of content.

[0115] Finally, the SCR inlet tracking NO at the current moment x The content is taken as the SCR inlet NO at the current moment xThe content signal is transmitted to the flue gas denitrification system as an input signal to obtain a time-delay prediction result.

[0116] Specifically, in the real-time operation of the model, the current moment is considered to be time t0, and the SCR inlet NO at this time is calculated. x The value of the content tracking curve at the current moment is ρ trace (t0), set the tracking NO x The content is the SCR inlet NO x The content signal is transmitted to the flue gas denitrification system as an input signal to obtain a time-delay prediction result.

[0117] The method for predicting the time lag of the SCR denitrification system in a thermal power plant of the present invention realizes the prediction of the time lag of the ammonia production system and can accurately predict the magnitude of the ammonia production time lag so as to coordinate the action of the ammonia injection system with the NO inlet of the SCR. x Content matching improves the prediction accuracy, saves costs for the overall operation of the SCR denitrification system, and provides suggestions for optimizing the control of the valve opening of the denitrification device.

[0118] like Figure 3 As shown, another aspect of the present invention provides a time lag prediction device 100 for a SCR denitration system of a thermal power plant, the prediction device 100 comprising:

[0119] The selection module 110 is used to select the SCR inlet NO according to the parameter-related selection principle within a set time period. x Concentration prediction parameters: The prediction parameters include at least but are not limited to SCR inlet flue gas temperature, SCR inlet flue gas oxygen content, load and total air volume.

[0120] Specifically, the selection module 110 selects prediction parameters based on the principle of maximum relevance and minimum redundancy, where:

[0121] The expression for the maximum correlation is:

[0122]

[0123]

[0124] Among them, maxD(S,c) is the maximum correlation of discrete variables, D is a categorical variable, S is a feature subset, c is a categorical variable, I is a feature variable, xi is the i-th feature, maxDF is the maximum correlation of continuous variables, and F(xi,c) is the F statistic;

[0125] The expression of minimum redundancy is:

[0126]

[0127]

[0128] Among them, minR(S) is the minimum redundancy of discrete variables, minRc is the minimum redundancy of continuous variables, and c(xi, xj) is the correlation function.

[0129] The maximum correlation and minimum redundancy are obtained based on mutual information. Mutual information can measure the correlation between two variables. The size of mutual information represents the size of the association between the two variables. By selecting a set number of features with the largest mutual information between the two variables from all features, we can get the set number of features that are most correlated between the two variables.

[0130] The expression of mutual information is:

[0131]

[0132] Among them, x is a variable, y is a variable, and I(x, y) is the mutual information.

[0133] The first establishing module 120 is used to establish the SCR inlet NO based on the predicted parameters. x Concentration prediction model.

[0134] Specifically, first, the first establishment module 120 obtains the prediction parameter data corresponding to each moment based on the prediction parameter; secondly, the first establishment module 120 obtains the time series data based on the prediction parameter data corresponding to each moment; finally, the first establishment module 120 establishes the SCR inlet NO based on the time series data. x Concentration prediction model.

[0135] The second establishing module 130 is used to establish a time-lag prediction model for the valve opening of the denitration device based on the time-lag shape difference rate method.

[0136] Specifically, the second establishment module 130 adopts the time-delay shape difference rate method to calculate the SCR inlet NO x The concentration and its rate of change are input, and the time-lag value is output. The prediction model of the time-lag value of the valve opening of the denitrification device is obtained by segmented fitting;

[0137] The time-delayed shape difference method is:

[0138]

[0139] Where T is the offset time δT, and the partial curve shape difference rate when the offset time is T, t1 and t2 represent the partial SCR inlet NO x The starting and ending time points of the concentration curve, ρ and C represent the SCR inlet NO x Concentration and ammonia regulating valve opening, k t (ρ) and kt (C) respectively represent SCR inlet NO x The backward change rate of concentration and ammonia regulating valve opening at time t.

[0140] The prediction module 140 is used to predict the SCR inlet NO x The concentration prediction model and the time lag prediction model are used to predict the time lag of the ammonia production system and obtain the prediction results.

[0141] First, through the SCR inlet NO x The concentration time series prediction model predicts the SCR inlet NO in a preset time period. x The content fluctuation data is used to predict the SCR inlet NO x Time series prediction curve of content.

[0142] Specifically, for the SCR inlet NO x The data of a certain moment (called time t0) in the content fluctuation data is obtained by using the SCR inlet NO x The content time series prediction model predicts the data for a period of time after this moment and obtains the time series prediction curve ρ(d), where d is the difference between any point on the time series prediction curve and the moment t0 in the time direction, that is, the time deviation of a certain point, in seconds.

[0143] Secondly, the SCR inlet NO x The content hysteresis prediction model for the SCR inlet NO x The time series prediction value at any time on the time series prediction curve of the content is predicted to obtain the SCR inlet NO x The time-hysteresis numerical curve corresponding to the content.

[0144] Specifically, the predicted value ρ(d0) at any time d0 on the time series prediction curve ρ(d) is calculated using the SCR inlet NO x The content time-lag prediction model obtains the SCR inlet NO at this time point. x The hysteresis prediction value D(d0) corresponding to the content is obtained by calculating the hysteresis prediction value of each point on the time series prediction curve ρ(d) and connecting the discrete hysteresis values ​​to obtain the hysteresis value curve D(d).

[0145] Here, according to the timing prediction curve and the time hysteresis value curve, the SCR inlet NO x Content tracking curve; among them,

[0146] The SCR inlet NO x The content tracking curve is:

[0147] ρ trace (t) = ρ(t + D(d)),

[0148] Among them, ρ(d) is the time series prediction curve, d is the difference between any point on the time series prediction curve and time t0 in the time direction, D(d) is the time hysteresis numerical curve, ρ trace (t) is the time deviation d corresponding to any tracking moment on the timing prediction curve ρ(d) of the SCR inlet NO x Predicted value of content.

[0149] Specifically, the tracking time t(d) is defined as the time corresponding to the time lag value corresponding to a certain time on the time series prediction model migrating to the negative direction of the time axis, that is, t(d) = t0-D(d). x Content trace (t) is the time deviation d corresponding to any tracking moment on the timing prediction curve ρ(d) of the SCR inlet NO x The content prediction value. For each point on the time series prediction curve, its tracking NO x The SCR inlet NO can be obtained by connecting the discrete results. x Content tracking curve:

[0150] ρ trace (t) = p(t + D(d)),

[0151] Among them, ρ(d) is the time series prediction curve, d is the difference between any point on the time series prediction curve and time t0 in the time direction, D(d) is the time hysteresis numerical curve, ρ trace (t) is the time deviation d corresponding to any tracking moment on the timing prediction curve ρ(d) of the SCR inlet NO x Predicted value of content.

[0152] Finally, the SCR inlet tracking NO at the current moment x The content is taken as the SCR inlet NO at the current moment x The content signal is transmitted to the flue gas denitrification system as an input signal to obtain a time-delay prediction result.

[0153] Specifically, in the real-time operation of the model, the current moment is considered to be time t0, and the SCR inlet NO at this time is calculated. x The value of the content tracking curve at the current moment is ρ trace (t0), set the tracking NO x The content is the SCR inlet NO x The content signal is transmitted to the flue gas denitrification system as an input signal, and the prediction module 140 obtains a time-delay prediction result.

[0154] like Figure 4 As shown, another aspect of the present invention provides an electronic device 200, comprising:

[0155] One or more processors 210, one or more storage units 220, one or more storage units 220 are used to store one or more programs, when the one or more programs are executed by the one or more processors 210, the one or more processors can implement the data recording method described above. The electronic device 200 also includes one or more input units 230 and one or more output units 240, etc. These components of the electronic device 200 are interconnected via a bus system 250 and / or other forms of connection mechanisms. It should be noted that Figure 3 The components and structures of the electronic device 200 shown are merely exemplary and non-limiting. The electronic device 200 may also have other components and structures as required.

[0156] The processor 210 may be a central processing unit (CPU) or other forms of processing units having data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device 200 to perform desired functions.

[0157] The storage unit 220 may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory (cache), etc. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor may run the program instructions to implement the client functions (implemented by the processor) and / or other desired functions in the embodiments of the present invention described below. Various applications and various data may also be stored in the computer-readable storage medium, such as various data used and / or generated by the application.

[0158] The input unit 230 may be a device used by a user to input instructions, and may include one or more of a keyboard, a mouse, a microphone, a touch button, a touch screen, and the like.

[0159] The output unit 240 may output various information (eg, images or sounds) to the outside (eg, a user), and may include one or more of a display, a speaker, and the like.

[0160] Another aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the data recording method described above can be implemented.

[0161] The computer-readable medium may be included in the apparatus, device, or system of the present invention, or may exist independently.

[0162] Among them, computer-readable storage media can be any tangible media that contains or stores a program, which can be an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or equipment. More specific examples include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, an optical fiber, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0163] The computer-readable storage medium may also include a data signal propagated in a baseband or as part of a carrier wave, which carries a computer-readable program code, specific examples of which include but are not limited to electromagnetic signals, optical signals, or any suitable combination thereof.

[0164] It is to be understood that the above embodiments are merely exemplary embodiments used to illustrate the principles of the present invention, but the present invention is not limited thereto. For those of ordinary skill in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A method for predicting the time lag of a thermal power plant SCR denitration system, characterized in that: The method comprises: During the set time period, select SCR inlet NO according to the parameter selection principle. x Concentration prediction parameters; Based on the prediction parameters, the SCR inlet NO is established. x Concentration prediction model, including: Based on the prediction parameters, obtaining corresponding prediction parameter data at each moment; Based on the prediction parameter data corresponding to each moment, obtaining time series data; Based on the time series data, the SCR inlet NO is established. x Concentration prediction model, including: Take the time series data as the SCR entry NO x The input data of the concentration prediction model, and in the process of establishing the prediction model, the hidden state of the prediction model is updated; wherein, The expression of the hidden state is: h t =f(h t-1 ,x t )=f(Uh t-1 +Wx t +b), Among them, U∈R G*G , W∈R G*k , b∈R G is the model parameter, G is the hidden layer dimension, k is the input dimension, and f is the tanh activation function; The SCR inlet NO is established based on the time series data. x After the concentration prediction model is obtained, the method further comprises: Based on the SCR inlet NO x The concentration prediction model is trained, wherein when the training residual is less than a preset threshold, based on the SCR inlet NO x The concentration prediction model is trained; Based on the time-lag shape difference rate method, a time-lag prediction model for the valve opening of the denitrification device is established, which specifically includes: The time-delay shape difference rate method is used to determine the SCR inlet NO x The concentration and its rate of change are input, and the time-lag value is output. The prediction model of the time-lag value of the valve opening of the denitrification device is obtained by segmented fitting; The time-delay shape error rate method is: Where T is the offset time δT, and the partial curve shape difference rate when the offset time is T, t1 and t2 represent the partial SCR inlet NO x The starting and ending time points of the concentration curve, ρ and C represent the SCR inlet NO x Concentration and ammonia regulating valve opening, k t (ρ) and k t (C) respectively represent SCR inlet NO x The backward change rate of concentration and ammonia regulating valve opening at time t; Based on the SCR inlet NO x The concentration prediction model and the time lag prediction model predict the time lag of the ammonia production system and obtain prediction results, specifically including: Through SCR inlet NO x The concentration time series prediction model predicts the SCR inlet NO in a preset time period. x The content fluctuation data is used to predict the SCR inlet NO x Time series prediction curve of content; Adopt SCR inlet NO x The content hysteresis prediction model for the SCR inlet NO x The time series prediction value at any time on the time series prediction curve of the content is predicted to obtain the SCR inlet NO x Time-hysteresis numerical curve corresponding to the content; According to the timing prediction curve and the time hysteresis value curve, the SCR inlet NO x Content tracking curve; among them, The SCR inlet NO x The content tracking curve is: r trace (t)=ρ(t+D(d)), Among them, ρ(d) is the time series prediction curve, d is the difference between any point on the time series prediction curve and time t0 in the time direction, D(d) is the time hysteresis numerical curve, ρ trace (t) is the time deviation d corresponding to any tracking moment on the timing prediction curve ρ(d) of the SCR inlet NO x Predicted content value; The SCR inlet tracking NO at the current moment x The content is the current SCR inlet NO x The content signal is transmitted to the flue gas denitrification system as an input signal to obtain a time-delay prediction result.

2. The method according to claim 1, characterized in that The SCR inlet NO is established based on the LSTM neural network algorithm x Concentration prediction model.

3. The method according to claim 1, characterized in that The parameter correlation selection principle adopts the principle of maximum correlation and minimum redundancy; wherein, The expression of the maximum correlation is: Among them, maxD(S,c) is the maximum correlation of discrete variables, D is a categorical variable, S is a feature subset, c is a categorical variable, I is a feature variable, xi is the i-th feature, maxDF is the maximum correlation of continuous variables, and F(xi,c) is the F statistic; The expression of the minimum redundancy is: Among them, minR(S) is the minimum redundancy of discrete variables, minRc is the minimum redundancy of continuous variables, and c(xi, xj) is the correlation function.

4. The method according to any one of claims 1 to 3, characterized in that: The SCR inlet NO x The concentration prediction parameters include SCR inlet flue gas temperature, SCR inlet flue gas oxygen content, load and total air volume.

5. A time lag prediction device for SCR denitration system of thermal power plant, characterized in that: The method for predicting the time lag of the SCR denitration system of a thermal power plant according to any one of claims 1 to 4, wherein the prediction device comprises: The selection module is used to select the SCR inlet NO according to the parameter-related selection principle within the set time period. x Concentration prediction parameters; The first establishing module is used to establish the SCR inlet NO based on the prediction parameter. x Concentration prediction model; The second establishment module is used to establish a time-lag prediction model for the valve opening of the denitration device based on the time-lag shape difference rate method; A prediction module is used to predict the SCR inlet NO based on the x The concentration prediction model and the time lag prediction model predict the time lag of the ammonia production system and obtain prediction results.

Images